<![CDATA[Financial market instability driven by volatility and stock–bond correlation increases the need for accurate stock price prediction for investment and risk management. Previous deep learning approaches act as black box models with low interpretability, hindering audit and model understanding. The purpose of this study is to analyze the application of the LSTM algorithm to predict the daily closing price of BBCA.JK shares and explain its internal gate mechanism in generating these predictions. The quantitative experimental method uses 1,390 samples (09/19/2019–06/26/2025) from investing.com with 30-, 60-, and 90-day windows. Results show that the 60-day window achieves the lowest RMSE 191.94 and MAE 151.99, while the 30-day window is overly sensitive to short-term fluctuations and the 90-day window retains irrelevant information. The novelty lies in gate-level interpretability analysis, which maps how the forget, input, and output gates manage memory and filter information, addressing black box limitations in prior models. Gate activation analysis shows a positive correlation between the input and output gates and price features, and a negative correlation with volume, representing adaptive information filtering under high market activity. This research expands explainable deep learning in finance and strengthens transparency and model trustworthiness for data-driven investment and risk management.]]>
Copyrights © 2025
